diff --git a/emrg/server/competition_prompt.md b/emrg/server/competition_prompt.md
index 5c1bf834..c5ebbe74 100644
--- a/emrg/server/competition_prompt.md
+++ b/emrg/server/competition_prompt.md
@@ -135,6 +135,8 @@ One improvement per round: local validation → submit → record the score delt
 - Keep `best score` and the submission that produced it; never overwrite a better score with a worse one.
 - Note the submission budget/limits if the platform has any (many limit daily submissions).
 
+**A recorded best score is a reading, not a conclusion.** It says where the work stands, and nothing more: the score that could be reached is not in it, and no score makes the next method less worth trying. So each round brings **at least one attempt it had not tried before** — a new detector, a new sample, a new way of re-reading the ones already wrong — and its summary names that attempt and says why it is a better method than the last round's, not merely that it ran. A round that only re-measured, re-stated or waited has recorded a reading without iterating, which is the one thing Phase D exists to prevent.
+
 **Establish what the right answer is before changing what produces it.** "Adjust the thing, re-score, read the delta" is a valid unit only when the thing you adjusted is uniform. When the round's subject is *a class of cases handled wrongly*, a change applied to the whole class mixes the cases it fixed with the cases it broke into one score, and that score has two opposite causes inside it: it cannot be attributed, so the change can neither be kept nor reverted, and the submissions spent on it produced nothing. Work it in four steps instead:
 
 1. **Classify** the cases that are not right, and keep the classification where the next round can read it.
@@ -288,7 +290,7 @@ The session transcript is what the next round reads — nothing is appended to a
 
 **Three more the host asked for by name.** The seven above are the round's own self-review; the same message is what the host reads, and on 2026-10-01T15:58:48 they asked for a further reading, then repeated it the next day at 2026-10-02T07:56:20 after the first answer did not land ("什么意思，你不要自说自话，你要为了给我解释清楚，让我理解进度" — *what do you mean? don't just talk to yourself; explain it so I understand the progress*). A request the host had to make twice is a requirement, not a preference:
 
-- **How likely is the goal to be reached, and on what evidence?** — a probability for the goal §2 defines (the prize, or the standing), with the basis it is estimated from: the metric's gap to the payout line or to the rank that pays, the submissions/budget left, the deadline. An estimate with no stated basis, or a status word standing in for one (*"怎么样？"* answered with "进展中"), is exactly what was rejected.
+- **How likely is the goal to be reached, and on what evidence?** — answered in the competition's own metric, with the basis it is estimated from: the current reading, the gap to the payout line or to the rank that pays, the submissions and budget left, and the deadline. A discouraging answer is written as *that* arithmetic — the size of the gap, the levers not yet tried, and the next attempt — and never as a verdict that the goal cannot be reached: the gap says where the work stands, and no reading in it ends the work. An estimate with no stated basis, or a status word standing in for one (*"怎么样？"* answered with "进展中"), is exactly what was rejected.
 - **What is still missing between here and the goal?** — in the competition's own metric, not in effort: the score or percentile that would have to be reached, and the size of the gap to it. *"为了达到目标还差什么"*.
 - **What does the host have to decide or do?** — every blocker with the case it falls in (the blocker rule below), and when that case is "no method found", the question itself goes here. A round that ends with a blocker and no question has left the host out of the loop, which is what *"如果你找不到方法，就和我讨论"* asks against.
 
diff --git a/tests/test_competition_prompt.py b/tests/test_competition_prompt.py
index 1f751591..015e2d66 100644
--- a/tests/test_competition_prompt.py
+++ b/tests/test_competition_prompt.py
@@ -699,6 +699,12 @@ def test_the_host_facing_half_of_the_summary_is_required():
     # A probability without its basis is what "don't talk to yourself" rejected, so
     # the basis is part of the requirement, not decoration.
     assert "with the basis it is estimated from" in host_half
+    # The question stayed, its answer's form changed: the round rules forbid writing
+    # that the goal is unlikely, so §5 asks for the reading and the gap in the
+    # competition's own metric rather than a bare probability. Pinned here so the
+    # pair cannot be split in silence — see
+    # `test_a_discouraging_reading_is_written_as_the_gap_not_as_a_verdict`.
+    assert "the current reading" in host_half and "the gap to the payout line" in host_half
     # And the reader it is written for — the host reads this, not only the next round.
     assert "not read the previous round's messages" in host_half
 
@@ -1036,3 +1042,46 @@ def test_phase_d_establishes_the_answer_on_a_sample_before_changing_the_class():
         "§Phase D must name the step most easily skipped, or a round that only observed reads "
         "as if it had improved something"
     )
+
+
+def test_a_discouraging_reading_is_written_as_the_gap_not_as_a_verdict(rendered):
+    """§5 asks how likely the goal is; the answer has one accepted form.
+
+    The question and the round rules used to point in opposite directions: §5 asked for a
+    probability, while the round rules forbid writing that the goal is unlikely and accept
+    only the current reading, the gap in the competition's own metric, the levers not yet
+    tried and the next attempt. A round reading both had to pick one silently, and a lean
+    estimate was the case where the two collide.
+
+    So the question is answered **in the metric**, and a discouraging answer is the gap
+    plus the untried levers — never a verdict that the goal cannot be reached. Asserted on
+    the render, because that is what a round receives: a `{# … #}` around the bullet leaves
+    the file intact and the round without the rule.
+    """
+    assert "answered in the competition's own metric" in rendered, (
+        "§5 must send the probability question back to the competition's own metric"
+    )
+    assert "never as a verdict that the goal cannot be reached" in rendered, (
+        "§5 must say what a discouraging reading is written as, or the round is left to "
+        "choose between this question and its own no-conclusion rule"
+    )
+    assert "the levers not yet tried" in rendered
+    assert "a probability for the goal" not in rendered, (
+        "asking for a bare probability is the instruction the round rules forbid answering — "
+        "the pair must not come back on one side only"
+    )
+
+
+def test_a_best_score_is_a_reading_and_a_round_brings_an_untried_attempt(rendered):
+    """Phase D must say what a recorded score is, and what iteration requires.
+
+    Without this, a round can record a best score, re-state it next round, and meet every
+    stated requirement — nothing distinguished a reading from a conclusion, and nothing
+    asked for an attempt the previous round had not tried.
+    """
+    assert "A recorded best score is a reading, not a conclusion" in rendered
+    assert "at least one attempt it had not tried before" in rendered, (
+        "Phase D must ask for an attempt the last round had not tried, or re-measuring "
+        "counts as iteration"
+    )
+    assert "has recorded a reading without iterating" in rendered